attacklm 0.3.3__tar.gz → 0.4.0__tar.gz

{attacklm-0.3.3 → attacklm-0.4.0}/.gitignore

@@ -127,6 +127,8 @@ dmypy.json
 # ----- Local notes (keep HANDOFF.md / TASKS.md tracked if you want) -----
 *.local.md
 notes.md
+HANDOFF.md
+TASKS.md
 
 # ============================================================================
 # GENERATED / REGENERABLE DATA — excluded (users generate these themselves)

{attacklm-0.3.3 → attacklm-0.4.0}/CHANGELOG.md

@@ -4,6 +4,62 @@ All notable changes to AttackLM are documented in this file. Versions follow [Se
 
 ---
 
+## [0.4.0] — 2026-06-22 — MoE-safe training, retention eval, and experience replay
+
+### Added
+
+- **Experience-replay / mixed-corpus mixer** (`scripts/replay_mixer.py`).
+  - Stratified mixing of one or more replay sources into any fine-tuning batch.
+  - New CLI flags on `attacklm-train-all`:
+    `--replay-source`, `--replay-ratio`, `--replay-max-examples`,
+    `--replay-stratify` / `--no-replay-stratify`, `--replay-domain-ratios`.
+  - Caches combined datasets under `data/datasets/combined/replay_<hash>.jsonl`.
+  - Records replay source composition in `state.json`.
+  - New `replay-general` source skeleton with starter samples for code,
+    conversation, factual, and reasoning domains.
+  - New acquisition script: `scripts/acquire_replay_general.py`.
+- **`attacklm-eval` retention suite** (`scripts/eval_retention.py`).
+  - Measures target-task gain vs. pretraining-domain retention.
+  - CLI entry point and 30 hermetic tests.
+- **Advanced training options** in `scripts/train_template.py`:
+  - `--use-dora`, `--loftq-init`, `--bf16`, `--fp16`, `--fp32`.
+  - `--use-rslora` / `--no-use-rslora`.
+  - `--target-modules` for explicit LoRA target selection.
+  - `--moe-safe-target` for Mixture-of-Experts models: bf16 only, no 4-bit
+    quantization, router/lm_head excluded from LoRA targets.
+- **Auto bf16 default** on Ampere/Ada/Hopper/Blackwell GPUs (compute capability >= 8.0),
+  with `--fp16` override for backward compatibility.
+
+### Changed
+
+- `scripts/train_all.py` now wires all new LoRA/DoRA/LoftQ/bf16/MoE-safe flags
+  through `build_train_cmd()` and integrates the replay mixer.
+- `scripts/train_template.py` records replay provenance in `state.json`.
+- `src/attacklm/cli.py` adds `attacklm-train-lora` and `attacklm-eval` entry points.
+- `pyproject.toml` registers `attacklm-eval` and `attacklm-train-lora` console scripts.
+
+### Fixed
+
+- `tests/test_thinking_models.py` updated to match the deprecated/removed
+  thinking-model helpers; now tests only the surviving `strip_thinking()` logic.
+- `tests/test_balance_buckets.py` updated its data-dependent total assertion
+  to match the current manifest.
+
+---
+
+## [0.3.3] — 2026-06-11
+
+### Fixed
+
+- `pyproject.toml` dynamic version resolution.
+- Bad v0.3.2 publish artifact.
+
+## [0.3.2] — 2026-06-11
+
+### Added
+
+- PyPI trusted publishing workflow (`.github/workflows/release.yml`).
+
 ## [0.3.1] — 2026-06-11 — `attacklm-init` one-shot setup
 
 ### Added

{attacklm-0.3.3 → attacklm-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: attacklm
-Version: 0.3.3
+Version: 0.4.0
 Summary: QLoRA-fine-tuned Qwen2.5-Coder for offensive security and AI red-teaming
 Project-URL: Homepage, https://github.com/Veedubin/AttackLM
 Project-URL: Repository, https://github.com/Veedubin/AttackLM

attacklm-0.4.0/REPLAY_PLAN.md

@@ -0,0 +1,209 @@
+# AttackLM — Experience-Replay / Mixed-Corpus Plan
+
+> **Status:** draft plan — pending implementation  
+> **Date:** 2026-06-22  
+> **Goal:** add a small, license-clean general-domain replay corpus to AttackLM fine-tuning batches so we can mitigate catastrophic forgetting, especially when fine-tuning MoE models.
+
+---
+
+## Why this matters
+
+AttackLM's training data is almost entirely offensive-security / red-team / MITRE content. When fine-tuning a dense or MoE model on this narrow distribution, the model rapidly forgets general coding, reasoning, factual, and conversational knowledge. For MoE models the risk is worse: the router can collapse so that nearly all tokens are routed through a small set of "security" experts, starving the rest.
+
+Experience replay keeps a small percentage of general-domain data in every batch. It acts as a regularizer on the shared layers and the router, preserving base-model capabilities while still learning the target domain.
+
+---
+
+## What to add
+
+Create one or more **replay sources** inside the existing per-source layout. The mixer supports multiple sources so you can layer a general replay corpus on top of domain-specific replay corpora.
+
+```
+data/datasets/buckets/sources/
+├── _index.json
+├── ...existing 11 sources...
+├── replay-general/                 # NEW — broad general-domain replay
+│   ├── LICENSE.md
+│   ├── SOURCE.md
+│   └── base/
+│       └── replay/
+│           ├── data_code.jsonl          # general code
+│           ├── data_conversation.jsonl  # assistant-style dialogues
+│           ├── data_factual.jsonl       # factual / encyclopedic text
+│           └── data_reasoning.jsonl     # reasoning / task instructions
+└── replay-coding/                    # NEW — optional extra code replay
+    ├── LICENSE.md
+    ├── SOURCE.md
+    └── base/
+        └── replay/
+            └── data_code.jsonl
+```
+
+This mirrors the provenance pattern used by every other source: each record carries `source`, `source_uri`, `license`, `license_uri`, and `rights_contact` fields. A re-distributor can drop the whole directory if they don't want the mixed-license bundle.
+
+---
+
+## Recommended source mix (per 1,000 replay examples)
+
+| Domain                                | Examples | Proposed source(s)                            | License (verify before ingest) |
+| ------------------------------------- | -------- | --------------------------------------------- | ------------------------------ |
+| **Coding**                            | 300      | The Stack v2 — permissive split               | Apache-2.0 / MIT / BSD mix     |
+| **Conversations / instruction following** | 250      | OpenAssistant + Anthropic HH-RLHF harmless    | Apache-2.0 + MIT (verify)      |
+| **Factual / encyclopedic**            | 250      | SlimPajama sample                             | Apache-2.0 (verify)            |
+| **Reasoning / QA / tasks**            | 200      | Natural Instructions / FLAN public pool         | Apache-2.0 (verify)            |
+
+The exact counts are a starting point. The mixer should let the user override the ratio per domain if retention evaluation shows one area degrading faster than others.
+
+---
+
+## Sources in detail
+
+### 1. The Stack v2 (permissive split) — coding
+- **What:** a large corpus of permissively licensed source code.
+- **Use:** sample Python, C, JavaScript, shell, and Go snippets to keep coding ability.
+- **License:** HuggingFace distributes a pre-filtered "permissive" version; verify before use.
+- **Risk:** low if you use the permissive split and avoid GPL/AGPL/CC-NC code.
+
+### 2. OpenAssistant — conversations
+- **What:** human-generated, assistant-style conversations.
+- **Use:** preserve general chat / instruction-following tone and safety behaviors.
+- **License:** Apache-2.0 (verify current release).
+- **Risk:** low.
+
+### 3. Anthropic HH-RLHF (harmless subset) — conversations
+- **What:** harmless/helpful dialogue data.
+- **Use:** retain harmless, helpful assistant behavior.
+- **License:** MIT for the HH-RLHF release (verify).
+- **Risk:** low; use the harmless split, not the red-team/adversarial split.
+
+### 4. SlimPajama — factual / web / books
+- **What:** a cleaned, deduplicated subset of RedPajama.
+- **Use:** retain general world knowledge and fluent English.
+- **License:** Apache-2.0 (verify current release).
+- **Risk:** low.
+
+### 5. Natural Instructions / FLAN public pool — reasoning
+- **What:** a broad collection of NLP task instructions: QA, NLI, summarization, translation.
+- **Use:** preserve instruction-following and reasoning.
+- **License:** Apache-2.0 (verify).
+- **Risk:** low.
+
+### 6. Dolly 15k (optional fallback)
+- **What:** small, human-generated prompt-completion dataset.
+- **Use:** quick instruction-following replay if you don't want to download large corpora.
+- **License:** CC BY-SA 3.0 (verify). Requires share-alike attribution; acceptable for research-only use but complicates redistribution.
+- **Risk:** medium due to share-alike.
+
+### 7. Self-generated replay (fallback / supplement)
+- **What:** use the base model to generate 500–1,000 diverse prompt-response pairs across coding, reasoning, factual QA, and conversation.
+- **Use:** zero external-license risk; lets you target the exact domains where retention evaluation shows weakness.
+- **License:** AttackLM-generated. If generated with Qwen2.5-Coder, the project already tags LLM-generated output as GPL-3.0 (see `sources/llm-generated/`).
+- **Risk:** may amplify base-model biases; must filter out low-quality or unsafe outputs.
+
+---
+
+## What NOT to use
+
+Consistent with AttackLM's existing high-risk source exclusions, avoid:
+
+| Source / license class          | Why avoid                                                       |
+| ------------------------------- | --------------------------------------------------------------- |
+| GPL / AGPL code or datasets     | Viral copyleft, conflicts with MIT-licensed project             |
+| CC BY-NC or any non-commercial  | AttackLM is meant for research but redistribution becomes unclear |
+| OpenAI-generated data (Alpaca, ShareGPT) | Terms-of-Service ambiguity for redistributing model outputs   |
+| Leaked / copyright-unclear prompts (e.g. TheBigPromptLibrary) | Already moved to `archive/restricted-sources/`            |
+
+---
+
+## How to mix it into training
+
+Add two new flags to `scripts/train_all.py` and `scripts/train_template.py`:
+
+```bash
+attacklm-train-all --single-model \
+  --dataset base/ tools/metasploit/ \
+  --replay-source replay-general/ \
+  --replay-ratio 0.07
+```
+
+### Proposed flags
+
+| Flag                       | Default | Description                                                                   |
+| -------------------------- | ------- | ----------------------------------------------------------------------------- |
+| `--replay-source`          | `None`  | Path/alias to one or more replay sources. Repeatable.                       |
+| `--replay-ratio`           | `0.0`   | Fraction of each fine-tuning batch that should be replay examples.            |
+| `--replay-max-examples`    | `0`     | Cap total replay examples across all sources (0 = ratio × target size).       |
+| `--replay-stratify`        | `True`  | Keep the per-source / per-domain mix instead of sampling uniformly.         |
+| `--replay-domain-ratios`   | `None`  | Optional JSON/object override of domain weights (e.g. `{"code":0.3, ...}`).  |
+
+### Mixing algorithm
+
+1. Load the target dataset (e.g., 10,000 security examples).
+2. Compute replay budget: `replay_ratio * len(target)` capped by `--replay-max-examples`.
+3. For each configured `--replay-source`, discover all `*.jsonl` files under `<source>/base/replay/`.
+4. Group files by domain (file stem after `data_`) and source. Apply per-source/domain weights.
+5. Stratify the budget across sources/domains.
+6. Combine target + replay, shuffle with a fixed seed, and write the combined JSONL.
+7. Record replay composition in `state.json[dataset]["replay"]`.
+
+Example state snippet:
+
+```json
+{
+  "dataset": {
+    "target_examples": 10000,
+    "replay_examples": 700,
+    "replay_ratio": 0.07,
+    "replay_sources": {
+      "data_code.jsonl": 210,
+      "data_conversation.jsonl": 175,
+      "data_factual.jsonl": 175,
+      "data_reasoning.jsonl": 140
+    }
+  }
+}
+```
+
+---
+
+## MoE-specific notes
+
+- **Router entropy is the key metric.** If pretraining-domain perplexity stays flat but routing entropy drops, the model is over-specializing. Increase the conversation + factual replay ratio first.
+- **Do not fine-tune router layers.** This is already covered by `--moe-safe-target` in `train_template.py`; replay makes that restriction safer.
+- **Avoid 4-bit quantization for MoE.** Use `--moe-safe-target` (bf16, no BnB 4-bit) when running the MoE pilot.
+
+---
+
+## Evaluation checklist
+
+Use the new `attacklm-eval` retention suite to decide if replay is working:
+
+- [ ] Baseline `attacklm-eval` on the base model before fine-tuning.
+- [ ] Fine-tune **without** replay and re-run `attacklm-eval`.
+- [ ] Fine-tune **with** replay (5%) and compare perplexity + QA accuracy.
+- [ ] If pretraining perplexity rises by > 0.2 nats or QA accuracy drops by > 5%, increase `--replay-ratio`.
+- [ ] If target-task performance drops, reduce `--replay-ratio` or remove one domain at a time.
+
+---
+
+## Implementation tasks
+
+- [x] Create `data/datasets/buckets/sources/replay-general/` with `LICENSE.md` and `SOURCE.md`.
+- [x] Create `scripts/replay_mixer.py` — load replay sources, sample with stratification, return a combined JSONL path + composition stats.
+- [x] Add `--replay-source`, `--replay-ratio`, `--replay-max-examples`, `--replay-stratify`, `--replay-domain-ratios` to `scripts/train_all.py`.
+- [x] Integrate mixer call into `train_all.py` after `build_combined()` (single-model) and per-bucket (multi-agent) paths.
+- [x] Add replay composition to `state.json` and `summary.json`.
+- [x] Write pytest tests for `replay_mixer.py` (hermetic, temp JSONL).
+- [ ] Verify exact licenses for each proposed source (must be done before running `scripts/acquire_replay_general.py`).
+- [x] Populate `data/datasets/buckets/sources/replay-general/base/replay/` with starter sample data so the mixer has something to work with.
+- [ ] Replace starter samples with real upstream data using `scripts/acquire_replay_general.py`.
+- [ ] Update `AttackLM/CHANGELOG.md` and `README.md` quickstart.
+
+---
+
+## Open questions
+
+1. Should the mixer support per-bucket replay (different replay mix per MITRE tactic) or one global mix?
+2. Should we generate a self-replay sample automatically from the base model as a default fallback when no external replay source is configured?
+3. Should replay examples be held out from the eval split so retention eval sees only target-domain validation data?
+

attacklm-0.4.0/data/datasets/buckets/sources/replay-general/LICENSE.md

@@ -0,0 +1,43 @@
+# replay-general — License Notice
+
+This is a **mixed-license** planned source for general-domain replay data used
+to mitigate catastrophic forgetting during AttackLM fine-tuning.
+
+## Status
+
+**Planned / not yet populated.** No data files exist under this directory yet.
+
+## Expected Data Sources and Licenses
+
+Each domain file (`data_<domain>.jsonl`) will carry per-record provenance
+fields (`source`, `source_uri`, `license`, `license_uri`, `rights_contact`)
+as required by the AttackLM attribution system.
+
+Planned upstream sources and their expected licenses:
+
+| Domain | Planned Source | Expected License | Risk Level |
+| --- | --- | --- | --- |
+| `code` | The Stack v2 (permissive split) | Apache-2.0 / MIT / BSD mix | Low |
+| `conversation` | OpenAssistant OASST1 | Apache-2.0 | Low |
+| `conversation` | Anthropic HH-RLHF (harmless) | MIT | Low |
+| `factual` | SlimPajama (sampled) | Apache-2.0 | Low |
+| `reasoning` | FLAN / Natural Instructions (public pool) | Apache-2.0 | Low |
+
+**Important:** Each license must be **verified before ingestion**. Do not
+assume the licenses listed above are current or accurate — check the upstream
+repository's LICENSE file and any dataset card on HuggingFace.
+
+## Excluded Licenses
+
+Consistent with AttackLM's high-risk source exclusions:
+
+- **GPL / AGPL**: Viral copyleft; conflicts with MIT-licensed project
+- **CC BY-NC / non-commercial**: Redistribution terms unclear
+- **OpenAI-generated data** (Alpaca, ShareGPT): ToS ambiguity
+- **Copyright-unclear scrapes**: Moved to `archive/restricted-sources/`
+
+## Redistribute Notice
+
+If you redistribute this source directory, you **must** include this
+`LICENSE.md` file and honor the per-record license fields. Each record's
+`license` and `license_uri` fields take precedence over this summary.

attacklm-0.4.0/data/datasets/buckets/sources/replay-general/SOURCE.md

@@ -0,0 +1,76 @@
+# replay-general — Source Description
+
+## Identity
+
+- **Source name:** replay-general
+- **Source type:** General-domain replay corpus (experience replay)
+- **Risk level:** Low
+- **Upstream URL:** See domain-specific sources below
+- **Local path:** `data/datasets/buckets/sources/replay-general/`
+
+## Purpose
+
+This source provides general-domain training examples mixed into AttackLM
+fine-tuning batches to mitigate **catastrophic forgetting**. When fine-tuning
+on narrow offensive-security data, the model rapidly loses coding, reasoning,
+conversational, and factual capabilities. A small fraction of replay data
+(5–10%) acts as a regularizer on shared layers and MoE routers.
+
+## Planned Domains
+
+| File | Domain | Planned Upstream | Upstream URL |
+| --- | --- | --- | --- |
+| `data_code.jsonl` | General coding | The Stack v2 (permissive split) | https://huggingface.co/datasets/bigcode/the-stack-v2 |
+| `data_conversation.jsonl` | Assistant dialogues | OpenAssistant OASST1 + Anthropic HH-RLHF (harmless) | https://huggingface.co/datasets/OpenAssistant/oasst1 |
+| `data_factual.jsonl` | Encyclopedic / web text | SlimPajama (sampled) | https://huggingface.co/datasets/cerebras/SlimPajama-627B |
+| `data_reasoning.jsonl` | Instruction / QA / reasoning | FLAN / Natural Instructions (public pool) | https://huggingface.co/datasets/google/flan_t5_xxl |
+
+## Data Format
+
+Each JSONL record follows the AttackLM per-source provenance convention:
+
+```json
+{
+  "messages": [
+    {"role": "user", "content": "..."},
+    {"role": "assistant", "content": "..."}
+  ],
+  "source": "replay-general/code",
+  "source_uri": "https://huggingface.co/datasets/bigcode/the-stack-v2",
+  "license": "Apache-2.0",
+  "license_uri": "https://www.apache.org/licenses/LICENSE-2.0",
+  "rights_contact": "legal@example.com",
+  "domain": "code"
+}
+```
+
+## Ingestion Notes
+
+- **No data files exist yet.** This is a skeleton for planned ingestion.
+- Each domain file must be independently downloadable and license-verified
+  before population.
+- Target: ~1,000 total examples (300 code, 250 conversation, 250 factual,
+  200 reasoning) for a 7% replay ratio on a 15K security dataset.
+- The `scripts/replay_mixer.py` tool handles discovery, stratified sampling,
+  and mixing into training batches.
+
+## Integration
+
+Use with `attacklm-train-all`:
+
+```bash
+attacklm-train-all --single-model \
+  --dataset base/ \
+  --replay-source replay-general/ \
+  --replay-ratio 0.07
+```
+
+Or with custom domain weights:
+
+```bash
+attacklm-train-all --single-model \
+  --dataset base/ \
+  --replay-source replay-general/ \
+  --replay-ratio 0.07 \
+  --replay-domain-ratios '{"code":0.3,"conversation":0.25,"factual":0.25,"reasoning":0.2}'
+```

attacklm-0.4.0/data/datasets/buckets/sources/replay-general/base/replay/data_code.jsonl

@@ -0,0 +1,3 @@
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: code."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "def greet(name: str) -> str:\n    return f\"Hello, {name}!\"\n"}], "source": "the-stack-v2-permissive", "source_uri": "https://huggingface.co/datasets/bigcode/the-stack-v2-0.3", "license": "Apache-2.0 AND MIT AND BSD-3-Clause", "license_uri": "https://huggingface.co/datasets/bigcode/the-stack-v2-0.3", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "the-stack-v2-permissive", "domain": "code"}
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: code."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "import json\n\nwith open('data.json') as f:\n    records = json.load(f)\nprint(len(records))\n"}], "source": "the-stack-v2-permissive", "source_uri": "https://huggingface.co/datasets/bigcode/the-stack-v2-0.3", "license": "Apache-2.0 AND MIT AND BSD-3-Clause", "license_uri": "https://huggingface.co/datasets/bigcode/the-stack-v2-0.3", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "the-stack-v2-permissive", "domain": "code"}
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: code."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "// Sum an array of integers\nconst sum = (arr) => arr.reduce((a, b) => a + b, 0);\n"}], "source": "the-stack-v2-permissive", "source_uri": "https://huggingface.co/datasets/bigcode/the-stack-v2-0.3", "license": "Apache-2.0 AND MIT AND BSD-3-Clause", "license_uri": "https://huggingface.co/datasets/bigcode/the-stack-v2-0.3", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "the-stack-v2-permissive", "domain": "code"}

attacklm-0.4.0/data/datasets/buckets/sources/replay-general/base/replay/data_conversation.jsonl

@@ -0,0 +1,3 @@
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: conversation."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "Hello! How can I help you today? I'm happy to answer questions, help with writing, explain concepts, or assist with coding."}], "source": "OpenAssistant", "source_uri": "https://huggingface.co/datasets/OpenAssistant/oasst1", "license": "Apache-2.0", "license_uri": "https://www.apache.org/licenses/LICENSE-2.0", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "OpenAssistant", "domain": "conversation"}
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: conversation."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "To bake sourdough bread, mix flour, water, salt, and starter, then let it ferment overnight before shaping and baking in a hot Dutch oven."}], "source": "OpenAssistant", "source_uri": "https://huggingface.co/datasets/OpenAssistant/oasst1", "license": "Apache-2.0", "license_uri": "https://www.apache.org/licenses/LICENSE-2.0", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "OpenAssistant", "domain": "conversation"}
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: conversation."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "A good sleep routine includes a consistent bedtime, limited screen exposure an hour before bed, and a cool, dark room."}], "source": "OpenAssistant", "source_uri": "https://huggingface.co/datasets/OpenAssistant/oasst1", "license": "Apache-2.0", "license_uri": "https://www.apache.org/licenses/LICENSE-2.0", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "OpenAssistant", "domain": "conversation"}

attacklm-0.4.0/data/datasets/buckets/sources/replay-general/base/replay/data_factual.jsonl

@@ -0,0 +1,3 @@
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: factual."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "The Amazon rainforest is the largest tropical rainforest in the world, covering approximately 5.5 million square kilometers across nine countries."}], "source": "SlimPajama", "source_uri": "https://huggingface.co/datasets/cerebras/SlimPajama-627B", "license": "Apache-2.0", "license_uri": "https://www.apache.org/licenses/LICENSE-2.0", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "SlimPajama", "domain": "factual"}
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: factual."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "Photosynthesis is the process by which plants use sunlight, water, and carbon dioxide to produce glucose and oxygen."}], "source": "SlimPajama", "source_uri": "https://huggingface.co/datasets/cerebras/SlimPajama-627B", "license": "Apache-2.0", "license_uri": "https://www.apache.org/licenses/LICENSE-2.0", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "SlimPajama", "domain": "factual"}
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: factual."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "The speed of light in a vacuum is approximately 299,792 kilometers per second."}], "source": "SlimPajama", "source_uri": "https://huggingface.co/datasets/cerebras/SlimPajama-627B", "license": "Apache-2.0", "license_uri": "https://www.apache.org/licenses/LICENSE-2.0", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "SlimPajama", "domain": "factual"}

attacklm-0.4.0/data/datasets/buckets/sources/replay-general/base/replay/data_reasoning.jsonl

@@ -0,0 +1,3 @@
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: reasoning."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "Given that all squares are rectangles and not all rectangles are squares, a shape with four equal sides and four right angles must be a square, which is also a rectangle."}], "source": "natural_instructions", "source_uri": "https://huggingface.co/datasets/Muennighoff/natural-instructions", "license": "Apache-2.0", "license_uri": "https://www.apache.org/licenses/LICENSE-2.0", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "natural_instructions", "domain": "reasoning"}
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: reasoning."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "To determine the cheaper option per liter, divide the total price by the volume. If a 2-liter bottle costs $3 and a 1.5-liter bottle costs $2.50, the 2-liter option costs $1.50/L while the 1.5-liter option costs about $1.67/L."}], "source": "natural_instructions", "source_uri": "https://huggingface.co/datasets/Muennighoff/natural-instructions", "license": "Apache-2.0", "license_uri": "https://www.apache.org/licenses/LICENSE-2.0", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "natural_instructions", "domain": "reasoning"}
+{"messages": [{"role": "system", "content": "You are a helpful assistant. Domain: reasoning."}, {"role": "user", "content": "Continue or summarize the following."}, {"role": "assistant", "content": "If it rains, the picnic will be canceled. It is raining. Therefore, the picnic will be canceled."}], "source": "natural_instructions", "source_uri": "https://huggingface.co/datasets/Muennighoff/natural-instructions", "license": "Apache-2.0", "license_uri": "https://www.apache.org/licenses/LICENSE-2.0", "rights_contact": "see data/REMOVAL.md", "attribution_required": true, "derived_from": "natural_instructions", "domain": "reasoning"}

{attacklm-0.3.3 → attacklm-0.4.0}/pyproject.toml

@@ -213,6 +213,8 @@ attacklm-clone = "attacklm.cli:main_clone"
 attacklm-init = "attacklm.cli:main_init"
 attacklm-balance = "attacklm.cli:main_balance"
 attacklm-build = "attacklm.cli:main_build"
+attacklm-train-lora = "attacklm.cli:main_train_lora"
+attacklm-eval = "attacklm.cli:main_eval"
 
 # ── Hatch build configuration ─────────────────────────────────────────────────