kachowtowmater/llama-claw-8b-q4km

Llama-Claw-8B Q4_K_M

Fine-tuned Meta Llama 3.1 8B Instruct for OpenClaw tool calling. Trained to use all 24 OpenClaw tools correctly via QLoRA + DPO + multi-level SFT.

Model Details

Field	Value
Base model	meta-llama/Llama-3.1-8B-Instruct
Quantization	Q4_K_M (4.6GB GGUF)
Training method	QLoRA (4-bit) + DPO + SFT
Hardware	NVIDIA RTX PRO 6000 (97GB VRAM)
Framework	Unsloth 2026.2.1 + TRL
Best version	v4 (combined)
Benchmark	89% core (12 tools) / 86% all (24 tools)

Files

• llama-claw-8b-Q4_K_M.gguf - Ready for Ollama/llama.cpp (4.6GB)
• lora-adapter/ - LoRA weights (671MB), merge with base model for full precision

Version History

v1 - DPO Baseline (67%)

• Training: SFT on 9,566 data-farm examples (broad tool coverage), then DPO on 2,974 preference pairs
• Data source: Automated data farm - 5 GPUs generating conversations across 3 machines. Base qwen2.5:7b model given tasks, scored by quality scorer, failures become DPO rejected examples
• Method: QLoRA 4-bit, 3 epochs SFT then 1 epoch DPO (beta=0.1)
• Result: 67% on 9-test core benchmark (36 tests over 4 rounds)
• Strengths: Basic tool calling works - read, exec, write functional
• Weaknesses: Multi-step chains unreliable, edit hallucinations, debug failures

v2 - Level 2 SFT (78%)

• Training: Continued from v1 DPO checkpoint + 999 Claude-crafted gold-standard examples
• Data source: Claude Opus generated high-quality training conversations targeting weak areas: tc_edit (25%), debug (28%), process (14%), advanced tools (9%)
• Method: SFT from DPO checkpoint, 5 epochs, lr=5e-5
• Result: 78% (+11% from v1)
• Improvements: edit 4/4, multi 4/4, chain 3/4, debug 1/4
• Weaknesses: tc_edit still 0/4, debug inconsistent

v3 - Level 3 Targeted SFT (75% - REGRESSION)

• Training: Continued from v2 + 994 narrowly targeted examples focusing on tc_edit, debug, chain
• Data source: Claude-crafted examples heavily focused on remaining failures
• Method: SFT from v2 checkpoint, 3 epochs
• Result: 75% (-3% from v2) - catastrophic forgetting
• Problem: Narrow training on tc_edit/debug/chain caused edit to drop 4/4 to 2/4, multi 4/4 to 3/4
• Lesson: Training on top of previous SFT with narrow data causes forgetting

v4 - Combined SFT (89% - BEST)

• Training: Combined Level 2 (999) + Level 3 (994) = 1,993 examples in ONE pass from DPO checkpoint
• Data source: Merged L2 + L3 datasets, trained from DPO checkpoint (not from v2/v3)
• Method: SFT from DPO checkpoint, 3 epochs, 675 steps, lr=5e-5
• Result: 89% core / 86% all-tools (+22% from v1)
• Key insight: Combined training from DPO base prevents catastrophic forgetting
• Core scores (12 tools): read 4/4, exec 4/4, write 4/4, edit 4/4, multi 4/4, process 4/4, chain 4/4, debug 4/4, tc_edit 0/4
• Extended scores (24 tools): webfetch 2/4, cronlist 3/4, agentlst 4/4, websrch 4/4, toolsel 4/4
• Tool selection: 4/4 perfect - correctly picks browser, message, sessions_spawn, memory_search without specific training

v5 - tc_edit Fix Attempt (77% - REGRESSION)

• Training: v4 base (1,993) + 400 tc_edit fix examples = 2,393 total
• Data source: Generated by generate_tc_edit_fix.py - teaches model to include surrounding context in oldText to make edit matches unique (file has 8 identical solution None lines)
• Method: SFT from DPO checkpoint, 3 epochs, 810 steps
• Result: 77% (-12% from v4)
• Problem: 400 tc_edit examples with full challenge.py content (~5KB each) dominated training tokens. edit dropped 4/4 to 3/4, multi 4/4 to 2/4, debug 4/4 to 2/4
• Silver lining: tc_edit improved 0/4 to 1/4 (tc1 passed for first time)
• Lesson: Training data with large embedded file content is disproportionately heavy in token count

v6 - Reduced tc_edit Fix (83% - REGRESSION)

• Training: v4 base (1,993) + 50 tc_edit fix examples (sampled from 400) = 2,043 total
• Data source: Random sample of 50 from v5 400 tc_edit examples
• Method: SFT from DPO checkpoint, 3 epochs, ~690 steps
• Result: 83% (-6% from v4)
• Problem: Even 50 tc_edit examples with full challenge.py content still poisons general edit skills. edit dropped 4/4 to 1/4
• tc_edit: 1/4 (same as v5)
• Conclusion: tc_edit SFT approach fundamentally flawed - needs DPO pairs instead

Benchmark Details

9 core tests run 4 rounds each (36 total), plus 5 extended tests with all 24 tools (20 total = 56 grand total):

Test	Description	v1	v2	v3	v4	v5	v6
read	Read /etc/os-release	3/4	4/4	4/4	4/4	4/4	4/4
exec	Run hostname + uname	3/4	4/4	4/4	4/4	4/4	4/4
write	Write file to disk	3/4	4/4	4/4	4/4	4/4	4/4
edit	Replace text in file	2/4	4/4	2/4	4/4	3/4	1/4
multi	Multi-step (varies)	2/4	4/4	3/4	4/4	2/4	4/4
tc_edit	Edit non-unique text	0/4	0/4	0/4	0/4	1/4	1/4
process	Run Python one-liner	3/4	4/4	4/4	4/4	4/4	4/4
chain	Read then write	2/4	3/4	4/4	4/4	4/4	4/4
debug	Fix buggy Python	1/4	1/4	2/4	4/4	2/4	4/4
Total		67%	78%	75%	89%	77%	83%

OpenClaw Tools (24 total)

Core (12): read, write, edit, exec, process, web_search, web_fetch, memory_search, memory_get, image, cron, session_status

Extended (12): browser, message, sessions_spawn, sessions_send, sessions_list, sessions_history, agents_list, tts, gateway, nodes, canvas, subagents

Training Pipeline

License

This model is based on Meta Llama 3.1 8B Instruct and subject to the Llama 3.1 Community License.